Convolutional Neural Networks for Multi-Class Image Classification

A Custom CNN Approach on Fashion-MNIST

Abstract

The given work is a design, implementation and evaluation of a tailored Convolutional Neural Network (CNN) that can be trained to provide multi-class image classification in ten different clothing categories based on the Fashion-MNIST benchmark data set (Xiao, Rasul, and Vollgraf, 2017). The data sets include 70,000 28x28 grayscale 28x28 pixel images (60,000 training data and 10,000 test data). In the proposed CNN architecture, three convolutional-pooling blocks each containing a filter depth of 32, 64 and 128 are used, then finally a fully connected classifier with a softmax output layer to generate class probabilities in the ten categories. The data preprocessing steps involved pixel value rescaling in [0, 1] range, reshaping tensors to meet the Conv2D layer requirements, stratified train/validation/test splitting, and sparse integer label encoding with the sparse categorical cross-entropy loss. The Adam optimizer (Kingma and Ba, 2015) was used, with validation-based callbacks, i.e. Early Stopping and ReduceLROnPlateau, to reduce overfitting and guarantee generalizable behavior. After testing on the held-out test set, the final model had test accuracy of about 92% and test loss of about 23 percent indicating high generalization to unknown data. The confusion analysis has shown that classification errors were clustered across the visually similar category of upper-body garments, in particular, shirt, T-shirt, coat, and pullover, which is also very common in the Fashion-MNIST literature and is also primarily due to the low pixel density of the dataset (Xiao et al., 2017). This conclusion is supported by the fact that a small, purposely designed CNN architecture is an effective and computationally efficient solution to this benchmark classification problem, and that the performance gaps still exist largely due to natural limits of the data sets, and not due to architecture weaknesses.